Skip to main content

METAL: A Meta-Learning Assistant for Providing User Support in Machine Learning and Data Mining

This project aims at the development of methods and tools for providing support to users of machine learning (ML) and data mining (DM) technology. While the interest in such technology, particularly in the area of classification and prediction, is growing rapidly in industry and commerce, and a number of data mining tools are already available, such tools are still of limited use to end-users who are not experts in machine learning. This is due, in part, to the fact that ML systems are non-trivial and their number keeps increasing. As a result, users of ML/DM technology are faced with two major problems: model selection and method combination, i.e., (a) selecting the best/most suitable model/algorithm to use on a given application, and (b) combining or integrating this with useful and effective transformations of the data. Traditionally, these problems are resolved by trial-and-error or through consultation of experts. The first solution is time consuming and unreliable. The second solution is expensive and biased by the experts' own prejudices and preferences, rather than scientific systematicity. Clearly, neither solution is completely satisfactory for the non-expert end-users who wish to access a much-needed technology. Automatic and systematic guidance is required.

In reaction to this need, the central goal of METAL is to develop a prototype assistant system that supports users with model selection and method combination, and guides them through the space of experiments. The system will combine prior meta-knowledge with meta-level learning. For each constituent, known techniques will be consolidated and original ones developed to cope with novel learning situations and applications. The assistant's meta-knowledge base, which integrates expert-given meta-information and the results of meta-learning, provides key information about past usage of ML systems. This knowledge describes the conditions under which operations carried out in the past have succeeded or failed. The proposed system is to exploit this information when providing guidance to the user. Such guidance will not be restricted to selection of an appropriate method only, but will also suggest data transformation steps that are often crucial to obtain good results. Furthermore, the system's guidance will not be limited to a single choice. It may consist of a set of promising operations that the user may readily explore in an orderly fashion. The system will extend its meta-knowledge base dynamically, as it is used, and hence have the capability to adapt to specific environments.

The expected effects, and the criteria by which the success of the project will be measured, are improved utility of data mining tools and in particular a significant savings in experimentation time. METAL is not only at the forefront of science, it also provides a promising solution to many practical problems. Results of the project will be reported in a book and the prototype will be accompanied by a technical manual for use by developers from the European software industry who would decide to implement it as a commercial tool.

The project's home page is here.

Staff and Students

Christophe Giraud-Carrier, Peter Flach, YongHong Peng, Jim Farrand, Hilan Bensusan


DaimlerChrysler (Germany), Dialogis (Germany), OEFAI (Austria), LIACC (Portugal), University of Geneva (Switzerland)


This research is supported by the European Union (ESPRIT Reactive LTR 26.357).